Why do I have to provide an items.id column to the group by clause? - postgresql

Question

I want to return unique items based on condition, sorted by price asc. My query fails because Postgres wants items.id to be present in the group by clause. If it's included the query returns everything matching the where clause, which is not what I want. Why do I need to include the column?
select items.*
from items
where product_id = 1 and items.status = 'in_stock'
group by condition /* , items.id returns everything */
order by items.price asc
| id | condition | price |
--------------------------
| 1 | new | 9 |
| 2 | good | 5 |
| 3 | good | 3 |
I only want items with ids 1 and 3.
Update: Here's a fiddle using the answer below, which still produces the error:
http://sqlfiddle.com/#!1/33786/2

Answer 1

The problem is that PostgreSQL has no way of knowing which items records you want to take values from; that is, it can't tell that you want this:
| id | condition | price |
--------------------------
| 1 | new | 9 |
| 3 | good | 3 |
and not this:
| id | condition | price |
--------------------------
| 1 | new | 9 |
| 2 | good | 5 |
To fix this, you need to use some sort of aggregation function, such as MAX:
SELECT MAX(id) AS id,
condition,
MAX(price) AS price
FROM items
WHERE product_id = 1
AND status = 'in_stock'
GROUP BY condition
ORDER BY price ASC
which gives:
| id | condition | price |
--------------------------
| 1 | new | 9 |
| 3 | good | 5 |
(This restriction is part of the SQL standard, and most DBMSes enforce it. One exception is MySQL, which allows your query, but with the caveat that "The server is free to choose any value from each group, so unless they are the same, the values chosen are indeterminate" [link].)

Answer 2

SQL Fiddle
select *
from (
select distinct on (cond)
id, cond, price
from items
where product_id = 1 and items.status = 'in_stock'
order by cond, price
) s
order by price

Answer 3

The SQL standard requires this behaviour, though some databases like MySQL ignore it and instead return unpredictable results.
If there's more than one row for "cond = good" and you ask for the "id" of the row where "cond = good", which row should the database give you? The row with id = 3, or id = 2? How should it know which to pick? MySQL picks an arbitrary row if there are multiple candidates, but this isn't allowed by the standard.
In your case you seem to want to pick the lowest-price row for each condition.
PostgreSQL provides an extension, DISTINCT ON ..., to help with this. Clodaldo has demonstrated this in his answer, so I won't repeat that here. Using DISTINCT ON will be much more efficient than the example below.
The SQL-standard way would be to use a window to rank the results, then filter on the ranked data. Unfortunately this is pretty inefficient as it requires all rows that match the inner where clause to be collected and sorted.
SELECT *
FROM (
SELECT *, dense_rank() OVER w AS itemrank
FROM items
WHERE product_id = 1 AND items.status = 'in_stock'
WINDOW w AS (PARTITION BY cond ORDER BY price ASC)
) ranked_items
WHERE itemrank = 1;
(http://sqlfiddle.com/#!1/33786/19)
Another SQL-standard way is to use an aggregation subquery to find the min prices for each category then display all rows with the min price:
SELECT *
FROM items INNER JOIN (
SELECT cond, min(price) AS minprice
FROM items
WHERE product_id = 1 AND items.status = 'in_stock'
GROUP BY cond
) minprices(cond, price)
ON (items.price = minprices.price AND items.cond = minprices.cond)
ORDER BY items.price;
Unlike the DISTINCT ON version, though, this will display multiple entries if the lowest priced item has more than one entry with the same cond and price.
So.. you should really use the DISTINCT ON approach, but you need to understand it. Start with the PostgreSQL documentation here.
On a side note, newer PostgreSQL versions allow you to refer to any column of a table whose primary key you've listed in GROUP BY; they identify the functional dependency of the other columns on the primary key. So you don't have to aggregate other cols if you've mentioned the PK in newer versions. That's what the standard requires, but older versions weren't smart enough to figure it out and required all columns to be listed explicitly.
That's what people who ask this question usually want to know, but doesn't apply strictly to your question since it turns out you're trying to use GROUP BY to filter rows.